#研究分享#【社会机器人的新应用:拯救本地语言】

#研究分享#【社会机器人的新应用:拯救本地语言】世界所有口语语言中有43%受到威胁,每14天又有一种语言死亡。澳大利亚的语言动态研究中心(CoEDL)与Google合作,通过人工智能来建立语言模型,迅速学习和翻译这些本地濒危语言。同时将这些模型应用在机器人Opie中,旨在教导孩子们了解澳大利亚土着语言。 Opie是一款简单,低成本的机器人,由木头,一些平板电脑,扬声器,移动路由器和Raspberry Pi电脑制成。

The social robot that could help save indigenous languages

Jietu20180610-000730

Hundreds of indigenous Australian languages risk being lost as the number of people taught to speak them dwindles, but a small robot named Opie and a team of dedicated researchers could be part of the key to preservation.

Researchers at the ARC Centre of Excellence for the Dynamics of Language (CoEDL) are in a race against time to preserve or revive indigenous languages, only 140 of which out of more than 300 are still spoken today. Only 18 of them are still taught to children.

Recognising that it would take almost two million hours for human linguists to transcribe and process the 40,000 hours of recordings that CoEDL collected in partnership with indigenous communities, researchers Ben Foley and Professor Janet Wiles have been looking into artificial intelligence technology for ways to increase efficiency.

Last week, CoEDL and Google announced a partnership to transcribe indigenous languages and build AI models that could support linguistic work. So far they've built initial models for six Australian indigenous languages — Bininj Kunwok, Kriol, Mangarayi, Nakkara, Pitjantjatjara, Warlpiri and Wubuy — plus a further five languages spoken in the Asia Pacific.

Google offers an open-source AI platform called TensorFlow, which organisations can use to build their own bespoke AI and machine learning systems.

Prof Wile says AI will never be able to do the linguist's job, which requires a lot of human interaction with communities, but it could allow linguists to get much richer, stronger data out of their recordings in less time.

"You can’t trust automatic speech-to-text, even in english, to get it perfect. So what we want to do is integrate the linguist's workflow," she says, suggesting that software could highlight and cross-reference important context in a transcript, or could even use machine learning to get better over time if it was integrated into the data collections systems through apps.

"Some languages have kinship terms that we don’t have in english. How does a linguist even know what questions to ask? ‘Do you have a different word for your sister’s child compared to your brothers child?’ But if we had apps on iPhones that actually had these relationships and could illicit that data automatically. So the person speaking the language would be directly working with the curating and recording of that data".

Google's Daan Van Esch, a product manager who works with the company's language teams to integrate dialects into products like Translate,  says that helping heritage languages cross the digital divide is critically important in the connected era.

"Since everyone is on social media these days, it actually means a lot to people for their language to exist in a digital space", he says. Since half of all online content is stored in English, each language added to Google's automatic translation ability opens up a wealth of information online for speakers.

"Whether we're talking about using AI to help transcribe a bunch of recording from the field, or eventually gathering dictionaries … it means a lot to these communities", says Mr Van Esch.

The CoEDL team's long-term ambition is to build language recognition and analysis software into their social robot, named Opie, who was designed to help teach endangered languages to children.

"You need young children, who are just coming to terms with their languages, to have access to digital resources. How do you do that? Enter robots", Prof Wiles says.

"We wanted to build robots that could go to remote communities, could be acceptable in cultural senses and in multiple different languages".

Opie is a low-cost and transportable robot made out of wood, a couple of tablets, a speaker, a mobile router, a Raspberry Pi computer and a regular USB charger hub. One tablet screen shows stories, games and lessons while the other shows a face that reacts to the child's actions. The physical and software design for each Opie is developed together with the community that will be using it, with most going to language centres or creches.

One activity on the Opie designed with the Ngukurr Language Centre is based on a memory card game the community liked playing on one of the centre's old computers. The instructions are spoken in Kriol, but the tablet houses four different heritage languages. When a card is flipped over that shows a grasshopper, for example, the word for grasshopper in the heritage language is played out loud. The software is modular so that new stories or even whole languages can be added.

While an application like this could easily be put on an app store for tablets and phones, Prof Wiles says the robot format is extremely important for teaching children aged two to five, the perfect age for learning languages.

"Social dynamics is really fast; 200 milliseconds is the time between when I say something and you nod or respond ... and most of our technology can't do that fast enough," she says.

"So we wanted to build our own robots that would actually have fast response. So that's where the eyes come in."

Prof Wiles and Mr Foley say that Australian indigenous languages, being very closely tied to indigenous cultures as well as the geographic locations in which they were spoken, contain a wealth of valuable cultural information that should be preserved.

Not only can ongoing preservation efforts protect endangered languages — with benefits including increased health outcomes for the remaining speakers — but Mr Foley says processing of artefacts and historical documents could even lead to the transcription and revitalisation of dialects that have already become extinct.


Comments are closed.



无觅相关文章插件